Advanced integer-only approximations for ARM TOSA architecture. Pushing the boundaries of transformer efficiency on edge hardware.
From problem identification to novel solutions - explore the key chapters of this thesis.
Why deploying transformers on edge devices requires integer-only softmax approximations.
Read moreARM's hardware specification: integer operations, TABLE lookup, and NPU constraints.
Learn moreInteger arithmetic fundamentals, homogeneous operations, and non-linear challenges.
Explore topicsComplete workflow for converting LLMs to integer arithmetic.
Learn moreAdaptive multi-table approach achieving 80,000× better precision than baseline.
Learn moreTesting error propagation on Qwen2 language models through transformer layers.
Learn moreWhich methods work best for different hardware constraints and accuracy requirements.
Learn moreComplete methodology, analysis, experiments, and conclusions.
View PDFTest real-world transformer inference or dive into theoretical mathematical analysis.
Real-world transformer inference with quantized attention. Observe how approximation errors propagate through actual language model layers.
Mathematical comparison of softmax approximation methods. Analyze error characteristics across different input ranges and visualize precision trade-offs.